基于分层卷积神经网络的皮肤镜图像分类方法

doi:10.11959/j.issn.2096-6652.202147

摘要/Abstract

摘要：

针对皮肤镜图像数量不充足以及各类疾病之间影像数据不平衡的问题，提出一种融合类加权交叉熵损失函数和分层卷积神经网络的皮肤镜图像分类方法。首先对皮肤镜图像进行色彩恒常化处理，消除环境光源噪声；然后构建基于ResNet 50的分层卷积神经网络，并在迁移学习的基础上分别构建二分类和多分类卷积神经网络模型，根据皮肤镜图像的数量特点设置类加权交叉熵损失函数。实验结果表明，该方法具有较好的分类效果，分类准确率达到了85.94%，与未改进的分类模型ResNet 50相比，测试准确率提高了5.752%。

关键词: ResNet50, 皮肤镜图像分类, 分层卷积神经网络, 过拟合

Abstract:

In order to solve the problem of insufficient number of dermoscopic image and the imbalance of image data among various diseases, a classification method of dermoscopic image based on class weighted cross entropy loss function and hierarchical convolution neural network was proposed.Firstly, the dermoscopic image was processed by color constancy to eliminate the ambient light noise.Then, the hierarchical convolution neural network based on ResNet 50 was constructed, and the two classification and multi classification convolution neural network models were constructed respectively, and the class weighted cross entropy loss function was set according to the quantitative characteristics of the dermatoscopic image.The experimental results show that the method achieves good classification effect, and the classification accuracy reaches 85.94%.Compared with the improved classification model ResNet 50, the test accuracy is improved by 5.752%.

Key words: ResNet 50, classification of dermoscopic image, hierarchical convolution neural network, over fitting

中图分类号:

TP391.4

邵虹,张鸣坤,崔文成. 基于分层卷积神经网络的皮肤镜图像分类方法[J]. 智能科学与技术学报, 2021, 3(4): 474-481.

Hong SHAO,Mingkun ZHANG,Wencheng CUI. Classification method of dermoscopic image based on hierarchical convolution neural network[J]. Chinese Journal of Intelligent Science and Technology, 2021, 3(4): 474-481.

图/表 15

图1

图2

图3

图4

表1

M分类概率分布"

类别	A₁	A₂	A₃	A₄	A₅	???	A_M
概率	$\frac{N_{1}}{N}$	$\frac{N_{2}}{N}$	$\frac{N_{3}}{N}$	$\frac{N_{4}}{N}$	$\frac{N_{5}}{N}$	???	$\frac{N_{M}}{N}$

表1

表2

表3

表4

图5

图6

表5

图7

图8

图9

表6

参考文献 14

[1]	HAY R J , JOHNS N E , WILLIAMS H C ,et al. The global burden of skin disease in 2010:an analysis of the prevalence and impact of skin conditions[J]. Journal of Investigative Dermatology, 2014,134(6): 1527-1534.
[2]	BINDER M , SCHWARZ M , WINKLER A ,et al. Epiluminescence microscopy:a useful tool for the diagnosis of pigmented skin lesions for formally trained dermatologists[J]. Archives of Dermatology, 1995,131(3): 286-291.
[3]	BALLERINI L , FISHER R B , ALDRIDGE B ,et al. A color and texture based hierarchical K-NN approach to the classification of non-melanoma skin lesions[J]. Color Medical Image Analysis, 2013,10(4): 63-86.
[4]	SCHAEFER G , KRAWCZYK B , CELEBI M E ,et al. An ensemble classification approach for melanoma diagnosis[J]. Memetic Computing, 2014,6(4): 233-240.
[5]	CELEBI M E , KINGRAVI H A , UDDIN B ,et al. A methodological approach to the classification of dermoscopy images[J]. Computerized Medical Imaging and Graphics, 2007,31(6): 362-373.
[6]	SUMITHRA R , SUHIL M , GURU D S . Segmentation and classification of skin lesions for disease diagnosis[J]. Procedia Computer Science, 2015,45: 76-85.
[7]	崔文成, 张鹏霞, 邵虹 . 基于深度可分离卷积网络的皮肤镜图像病灶分割方法[J]. 智能科学与技术学报, 2020,2(4): 385-393.
	CUI W C , ZHANG P X , SHAO H . Dermoscopic image lesion segmentation method based on deep separable convolutional network[J]. Chinese Journal of Intelligent Science and Technology, 2020,2(4): 385-393.
[8]	CODELLA N , CAI J J , ABEDINI M ,et al. Deep learning,sparse coding,and SVM for melanoma recognition in dermoscopy images[C]// Proceedings of the 6th International Workshop on Machine Learning in Medical Imaging.[S.l.:s.n.], 2015: 118-126.
[9]	KAWAHARA J , BEN TAIEB A , HAMARNEH G . Deep features to classify skin lesions[C]// Proceedings of 2016 IEEE 13th International Symposium on Biomedical Imaging. Piscataway:IEEE Press, 2016: 1397-1400.
[10]	POMPONIU V , NEJATI H , CHEUNG N M . Deepmole:deep neural networks for skin mole lesion classification[C]// Proceedings of the 23rd IEEE International Conference on Image Processing. Piscataway:IEEE Press, 2016: 2623-2627.
[11]	ESTEVA A , KUPREL B , NOVOA R A ,et al. Dermatologist-level classification of skin cancer with deep neural networks[J]. Nature, 2017,542(7639): 115-11811.
[12]	李航, 余镇, 倪东 ,等. 基于深度残差网络的皮肤镜图像黑色素瘤的识别[J]. 中国生物医学工程学报, 2018,37(3): 274-282.
	LI H , YU Z , NI D ,et al. Identification of melanoma using dermoscopic images based on deep residual networks[J]. Chinese Journal of Biomedical Engineering, 2018,37(3): 274-282.
[13]	田敏 . 基于迁移学习的医学文献内图像多标签分类[D]. 武汉:武汉大学, 2019.
	TIAN M . Image multi label classification within medical literature based on migration learning[D]. Wuhan:Wuhan University, 2019.
[14]	HE K M , ZHANG X Y , REN S Q ,et al. Deep residual learning for image recognition[C]// Proceedings of the 29th IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2016.

皮肤镜图像	数量/张	占比
基底细胞癌	514	5.13%
光化性角化病	327	3.27%
良性角化病	1 099	10.97%
皮肤纤维瘤	115	1.15%
黑色素瘤	1 113	11.11%
黑色素细胞痣	6 705	66.95%
血管病变	142	1.42%

名称	参数
处理器	I7-7700HQ
内存	8 GB
显卡	GTX 1050 Ti
实验平台	TensorFlow
实验框架	Keras
语言	Python

epoch	原始数据	色彩恒常化处理过的数据
1	467 s	493 s
2	461 s	323 s
3	461 s	335 s
4	463 s	341 s
5	462 s	326 s
6	462 s	326 s
7	462 s	327 s
8	462 s	325 s
9	462 s	325 s
10	463 s	326 s

epoch	0～10	11～20	21～30
训练准确率	72.185%	78.355%	80.164%
测试准确率	71.419%	77.362%	78.938%

性能指标	计算方式	值
召回率	TP/(TP+FN)	90.27%
精确率	TP/(TP+FP)	92.39%
准确率	(TP+TN)/(TP+FP+FN+TN)	85.94%